Study On Water Quality Evaluation And Prediction Method Based On Multi-source Data Fusion

As the key means of water condition judgment and change trend analysis,water quality evaluation and prediction play an indispensable role in solving the problem of water pollution.However,the traditional evaluation and prediction methods cannot make full use of the characteristics of multi-source,nonlinear and uncertain water quality monitoring data,and the accuracy of the results is low.In order to better utilize the complex nature of water quality data and improve the accuracy of evaluation and prediction,this thesis proposes a water quality evaluation and prediction model based on multi-source data fusion technology.The main research contents are as follows:Firstly,an evaluation model based on the combination of neural network and Dempster-Shafer(D-S)evidence theory was proposed to solve the problem that traditional water quality evaluation methods could not make full use of the complex characteristics of monitoring data.The model firstly adopts back propagation(BP)neural network,radial basis function(RBF)neural network and generalized regression neural network(GRNN)to preliminarily evaluate the water monitoring data,and then the evaluation results are input into the D-S evidence theory for fusion evaluation.The experimental results of monitoring section in southern Hebei show that the evaluation model can reduce the uncertainty and improve the accuracy of water quality evaluation.Secondly,in order to solve the paradox problem of the fusion of high conflict evidence in D-S evidence theory,a weighted evidence fusion method based on conflict representation is proposed.Combined with Pignistic probability distance and cosine similarity between evidences,the degree of conflict between evidences is jointly measured,then which is converted into evidence weight.The conflicting evidence is replaced by weighted average evidence,and then the fusion result is iteratively modified.The simulation results of numerical examples show that the improved method can solve the conflicts between evidences well,improve the accuracy of fusion results,and avoid the conflict between results and facts.Finally,in view of the shortcomings of the long short term memory(LSTM)neural network such as complicated parameter determination and low prediction accuracy,an improved LSTM neural network prediction model with random forest(RF),particle swarm optimization(PSO)and attention mechanism(AM)is proposed.The model selects input features based on random forest,determines the optimal value of the number of neurons in the hidden layer and the learning rate of the neural network through the particle swarm algorithm,and introduces the attention mechanism to improve the learning ability of the neural network for key features.Through the prediction and error analysis of the dissolved oxygen,the high precision and goodness of fit of the model are verified.